Music-CRN: an Efficient Content-Based Music Classification and Recommendation Network

For human beings, music is generally perceived, categorized, and enjoyed based on its attributes, such as rhythm, pitch, timbre, and harmony. In recent years, due to their high performances, content-based music classification and recommendation systems have attracted much attention from both the music industry and research community. However, on the one hand, deep music classification models are still very rare, and on the other hand, the collaborative filtering approach, which has the cold start problem, still dominates the music recommendation applications. In this paper, we propose Music-CRN (short for music classification and recommendation network), a simple yet effective model that facilitates music classification and recommendation with learning the audio content features of music. Specifically, to extract the content features of music, the audio is converted into spectrogram “images” by Fourier transformation. Music-CRN can be applied on the spectrograms as similar as natural images to effectively extract music content features. Additionally, we collect a new dataset containing nearly 200,000 music spectrogram slices. To the best of our knowledge, this is the first publicly available music spectrogram dataset, which is at https://github.com/YX-Mao/Music-spectrum-image-data. We compare Music-CRN to previous content-based music classification and recommendation models on the collected dataset. Experimental results show that Music-CRN achieves state-of-the-art performance on music classification and recommendation tasks, demonstrating its superiority over previous methods.